UtilityBar runs overnight across every client's data and your agency's method, then hands you a Control Center: what changed, what to do about it, and the brief already written, with the evidence attached.
Three engines read the avalanche guide. None name you. Two competitors are named.
Overnight topic cluster audit. Retaining walls lost 404 clicks since Sept 12.
Two can go to the client now. Two cite a retailer page that changed.
Drafted in each account manager's voice, checked against CRM and open deals.
"Low gravity pavers" prompts stopped reaching your guide. Images carry the answer.
Comparison blocks won 3 of 3. FAQ schema was null twice.
0 · Agency
Your agency. Everything below answers one question: why would this not produce slop, and why is it yours? Scroll. Name the node if you like.
How this is different
Most agent tools hand you a blank chat and a list of tools. This one starts from the SOW: scope, campaign and approved tactics are set up before you type a word, the crawl writes its own settings, and what comes next is already picked from what the client bought.
sc-domain:havelockhotsauce.com is shared with our Google Cloud account · 16 months of dataIt lives in your chat
No copy and paste, no "here is the context". The chat is the plain kind you already use. What is different is what it arrives with, who you can bring into it, and what it keeps.
Build the Q4 content plan from the gaps. Stay inside the SOW.
Scorch & Sons out-ranks you on 61 terms, mostly heat-level guides, and gets named where you don't. This plan closes the three biggest gaps first.
Move the heat-scale guide to October and drop the gift guide.
Done. v2 has seven posts and the heat-scale guide leads.
Can this be the first post?
Love it. One thing: we never say “burn”. It's “heat”.
Noted, and kept. Replaced three uses of “burn” in the plan and the draft.
Prove what moved
Every intervention is dated in the ledger, so the question "did that work?" is a comparison, not a debate. Search Console data is clustered into topics and sub‑topics; each cluster is colored by what happened to it after the change.
| Topic | Clicks (before) | Clicks (after) | ↓ Δ |
|---|---|---|---|
| Regolith Paver Sizing | 143 | 2,894 | +2,751 |
| Free regolith paver spec sheets | 50 | 1,043 | +993 |
| Paver sizing for 1/6 g | 33 | 629 | +596 |
| Regolith paver thickness charts | 28 | 491 | +463 |
| Sintered regolith vs. cast basalt | 9 | 134 | +125 |
PagesQueries | |||
| sintered regolith pavers | 2 | 40 | +38 |
| cast basalt paver moon | 3 | 37 | +34 |
| regolith paver vs basalt | 1 | 19 | +18 |
| Low‑Gravity Installation | 412 | 1,602 | +1,190 |
| Crater‑Rim Retaining Walls | 1,150 | 746 | −404 |
| Habitat Patio Layouts | 1,083 | 900 | −183 |
The same view exists per page and per query, and the intervention that caused the change is one click away in the ledger. Losses show up as plainly as wins; that is the point.
AI relationship intelligence
The same engine, pointed at your CRM instead of your content. It finds the people worth a note this week, works out why now, drafts it in the account manager's voice, and files it in their library.
Searching for up to 50 candidates. This usually takes a few minutes — you can leave this page and come back.
Subject Congrats on Northbeam
Dana, saw the news. Congratulations. If the snack segmentation is still on your wall, I'd love to hear what you'd build at Northbeam. I can send the tracker guide you grabbed in August with our notes on it…
Refine with feedback…Send this approachThe slop question
Every agency we talk to has tried AI content and got slop. So did we. UtilityBar does not write your articles. It decides, briefs, audits and recommends on connected data, and nothing reaches a person until it has passed a chain of gates built for that job. Here is the chain on the weekly opportunity flow you just scrolled through.
Signals from the CRM, the web and the graph. A job change with no history on the account does not become a card.
Rejects: weak signal, no historyRouted to the right account manager. Collision detection catches another lead already on the account, or an open deal the CRM did not surface.
Rejects: wrong owner, live dealOnly the data and the report that fit this person and this moment. Twenty‑five page visits over three years, not a dump of the record.
Rejects: padding, stale dataThe ask has to match the context. A congratulations is not a pitch. A renewal is not a cold intro.
Rejects: wrong move for the momentDrafted in the account manager's voice from their library, then checked against it. Theirs to send, or to fix.
Rejects: generic toneThat is one flow. Every recommendation engine in UtilityBar has its own chain. The gates differ. The rule that every step has one does not.
Everything an agent reads comes from the knowledge graph: crawls, analytics, keyword clusters, CRM, your briefs and SOWs, assembled by pipelines we hardened on the edge cases. Context is gated, so the model sees what matters for this decision and nothing else.
Every correction is labeled by cause and stage, then deployed where it belongs: the pipeline, the data, or the playbook. The same mistake is not available next week.
Every change is an experiment with a result. You cannot keep making the same call when the ledger shows it as a loss, and neither can the system.
You review where review earns its keep: preview the page change, expand the reasoning, zip through a series across pages, deploy. We have not seen that anywhere else.
It does not write net‑new articles, and we have not built that pipeline. Briefs, audits, recommendations and plans, yes. A research pipeline is in progress (goals → research → report → ideate → experiment → artifact → page) and it will ship behind the same gates, grounded in the same graph. If your problem is net‑new technical content in a thin category, we will tell you so on the call.
Give your clients a seat
Clients log into your workspace, under your name, and see only what is theirs: drafts to approve, opportunities to review, what moved since last week. They stop waiting on a monthly deck, and they stop wondering what you do.
Built and run for you
We build, host, secure and operate it. Your team works inside it and ships with clients. The chat does the work, the Control Center makes the work legible, and the graph remembers what happened.
Per‑client overview, activity feed, motions (recommended → active → measuring → completed), integration health and reports. Client‑role logins see only what's theirs.
A small set of agents that each finish one job well: onboard a site, research a market, recommend, publish, report. Each hands off to the next with context intact.
Crawls and crawl configs, Search Console and keyword data, topic clusters, cannibalization, HubSpot and LinkedIn enrichment, citation monitoring, link placements, WordPress drafts, edits and QA.
Sites, pages, keywords, people, companies, placements and experiments as connected data, described by XEO, our open ontology for how engines see, cite and serve content.
Every intervention is logged: what was done, for which site, with which tactic, when. Before/after analysis stops being a slide and becomes a query.
One instance per client. Your data, your API keys, your model choices. Read‑only, least‑privilege access to source systems. No commingling, no training on your data.
The cohort
We can run five agencies well at once. Two are on now. The three open seats are for agencies willing to put around 25 clients on the system while it is in beta, and to work closely with us while their method goes in.
Clients onboard in a day. An agency does not. Its method lives in the heads of its senior people, so loading it means SME interviews, data we have to earn access to, and weekly working sessions until the recommendations sound like you. That is engineer time we cannot fake, so we do not sell more of it than we have.
It is a lot to ask up front. In practice it arrives slowly as trust builds, and we plan for that.
Your data blended with ours, your method distilled into the engine. The advantage is yours and nobody else's.
Our standard, most reliable SEO and AEO flows on third‑party data, shaped by a one or two day workshop on your method.
In our first sales‑intelligence pilot, most of the cards account leads rejected were routing and ownership problems: wrong person, departed owner, an active deal the CRM didn't show. The model wasn't the constraint. The CRM was. So data hygiene and a written playbook are step one now, and adoption is a success criterion we instrument, not a hope.
Proof, not promises
We built UtilityBar for our own agency first. Two other agencies then funded pieces of it as paid proofs of concept, on their clients' data, with their teams in the loop. In production since January 2026. Numbers are from live tenants as of September 2026.
SEO and AI engineering agency in Philadelphia, since 2014. Our own client roster runs on UtilityBar every morning, and every feature on this page ran for our clients before it ran for anyone else.
Eight account leads, a year of weekly opportunity cards checked against CRM and project history, and a second year adding search visibility work. The cards and corrections counted below are theirs.
Funded the build of their link placement data for two business lines into the graph, their own MCP server, and the first agents on it. Their top 25 client sites are loaded, with keyword data, in a white‑labeled Control Center.
Questions we get
Not unless you decide it should. Opportunity cards are reviewed and sent by people. WordPress drafts land on staging with a preview link. Recommendations are approved before they become motions. How much review each step gets is your setting, not our limitation: teams start with everything gated and relax the steps the scoreboard has earned.
It does not produce articles, so it cannot produce slop articles. It produces decisions: briefs, audits and recommendations with plans, each grounded in the graph and passed through gates built for that flow. Where it drafts, such as an opportunity note or a comparison block, the draft is checked against your library and your rules, and you keep the send button. The gates section above walks through one chain.
Two routes. Blend is a beta seat: $20k a month for a six‑month build, then a run agreement at about 60% less. A narrower scope starts at $15k a month for three to four months. Base is our standard SEO and AEO flows on third‑party data after a short method workshop, from $5k a month. Both start with a discovery call where we tell you which fits, or that neither does.
Because loading an agency’s method is engineer time, not a signup form. SME interviews, data access we earn slowly, weekly working sessions. We can do that well for five agencies at once, two are on now, and we would rather say so than oversell it.
For relationship intelligence: CRM access (HubSpot today), a list of account leads and their accounts, and whatever project history you have, even a spreadsheet. For visibility operations: the site, Search Console access if you have it, and any citation or link‑placement data. We add crawls, keyword data and web research ourselves.
It usually is. We find out in week one, we tell you plainly, and cleanup becomes the first deliverable: duplicates, bounced contacts, missing owners, orphaned companies. Cards and recommendations are only as trustworthy as the records underneath them, so we don't skip this step.
Whatever fits the task and your policy. Each tenant runs with its own keys, so you choose the providers, you see the spend, and nothing is trained on your data. We swap models as the frontier moves without changing your workflows.
Yes. The Control Center is built as a tenant overlay: your name, your domain, your client roster. One agency already runs it under its own product name for its clients.
Assistants and agents are good at one‑off tasks you think to ask for. They will not decide which task matters for which client tomorrow morning, log what they changed, or remember that you rejected the same idea in March. UtilityBar ships prioritization, memory and accountability on your data, in your method, with people in the loop by design.
It is the layer those need. Fast typed decision models are only as good as the options, labels and thresholds you give them. The queue, your corrections and the experiment ledger are exactly that data: your decisions, recorded with their outcomes. We use the best model for each judgment and keep the workflow, thresholds and actions in code you can audit.
Get started
Philadelphia, PA · Working with clients across the US since 2014
